Datasets:

Aditvir
/

agentic-customer-support-india

license: mit
task_categories:
  - text-classification
  - text-generation
language:
  - en
tags:
  - customer-support
  - e-commerce
  - india
  - intent-classification
  - sentiment-analysis
  - retrieval-augmented-generation
size_categories:
  - n<1K

Agentic Customer Support Dataset (India E-Commerce)

An annotated customer-support query dataset for the Indian e-commerce domain, built to train and evaluate the Agentic Customer Support System (CPG 289, Thapar Institute of Engineering and Technology) — a multi-agent, hybrid-retrieval support architecture with a LoRA fine-tuned Llama 3 8B language model.

Dataset Summary

150 customer support queries, each labelled with intent (6 categories), sentiment (3 classes), difficulty (3 levels), and a grounded answer drawn from the project's real policy knowledge base — not an invented or hallucinated response.

Split	Entries
`train.jsonl`	105
`validation.jsonl`	22
`test.jsonl`	23

Data Fields

Field	Type	Description
`id`	int	Unique entry identifier
`query`	string	The customer's message
`intent`	string	One of: `refund`, `delay`, `cancellation`, `wrong_item`, `policy`, `general`
`sentiment`	string	One of: `positive`, `neutral`, `negative`
`difficulty`	string	One of: `easy`, `medium`, `hard`
`answer`	string	A grounded response, sourced from the project's real policy documents (see Provenance)

Example

{
  "id": 2,
  "query": "Do I get the shipping fee refunded too for order ORD10023?",
  "intent": "refund",
  "sentiment": "neutral",
  "difficulty": "medium",
  "answer": "Once a returned item is picked up and reaches Amazon's fulfilment centre, refund processing begins within 13 days of successful pickup. If no refund update appears after 13 days, customers should contact Customer Service for further assistance."
}

Provenance and Generation Methodology

Queries are generated from hand-written templates (5 phrasings per intent category, with lexical variety — not one query repeated with different labels), each populated with a real order ID drawn from the project's mock order dataset where relevant. Every answer is grounded in the project's actual knowledge base — a set of policy documents sourced and paraphrased from Amazon.in's real, published customer service pages (returns, refunds, cancellations, delayed delivery, and general support), not invented for this dataset.

Honest Scope Note

This dataset is 150 entries, programmatically generated, not the 1,000-entry, manually hand-curated dataset originally targeted at proposal stage. It is large enough to exercise the full label schema and produce a genuine train/validation/test split, and was used as-is to LoRA fine-tune the project's Llama 3 8B model. Scaling to 1,000 hand-verified entries remains documented future work — this release is not presented as that larger target, and reviewers should not assume hand-curation beyond the template and grounding process described above.

Intended Use

Training and evaluating small-model fine-tuning and retrieval-augmented generation systems for customer support in a multilingual, Indian e-commerce context. Not intended as a general-purpose sentiment or intent classification benchmark outside this domain.

Limitations

English-language queries only (the deployed system additionally supports Hindi via a separate translation layer not reflected in this dataset)
Template-based generation means limited syntactic diversity within each intent category
Order IDs and customer details are synthetic, not real customer data

Citation

If you use this dataset, please cite the associated capstone project:

Agentic Customer Support System (CPG 289)
Thapar Institute of Engineering and Technology, Patiala, 2026